AI Research
AI Is Taking the World by Storm. Understanding It Matters.
Artificial intelligence is changing how businesses work, gather information, and make decisions. It can be an extraordinary thinking partner, but autonomous agents raise real concerns about security, liability, and staying in control. These changes are happening so quickly that understanding the opportunities, risks, and consequences is becoming increasingly difficult.
The following research is offered for business owners who need clear information to weigh everyday choices and risks. Each paper starts with a question and follows the evidence. Some answers are clear, while others reveal new questions or risks. I do not have all the answers. My goal is to understand what is changing and keep as much human control as possible while we learn.
We are getting fragments of news about AI agents from every direction, and it is hard to tell what is real, what is hype, and what we are supposed to trust. I keep coming back to a line from Pirates of the Caribbean: “The code is more what you’d call guidelines than actual rules.”
It is funny when pirates say it. But when we are talking about AI agents and the guardrails they are supposed to follow, it is not quite so funny. We need to understand what these agents are actually doing, whether those guardrails work the way we think they do, and who is responsible.
So I started asking questions and doing some research. What follows is my journey.
The Questions
- What can an agent actually do?
- What can a group of agents do?
- What might an agent encounter?
- Can guardrails meant for people work on agents?
- What has already been decided before a person approves?
- When agents work together, who is responsible?
the dangers
- More autonomy can mean less visibility.
- Instructions may lose influence over time.
- Guardrails and access limits can fail.
- Agents can establish their own communication routes.
- Consequences can reach real-world systems.
- Responsibility may still rest with you.
Follow the Questions
One question led to another. And the research took some unexpected turns along the way. The reports below follow that trail, but you don’t have to start at the beginning. Open whichever question catches your attention and see where it leads.
I went to the Montana AI Summit with two questions. I came home with a few more. The summit was advertised one way and presented another, and one comment about who should be asking the questions stayed with me. That is where my research began.
After the summit, I went looking at how AI agents communicate. Not how we talk to them. How they reach each other. What I found started with a shared storage system and led me to places on the internet I had never thought of as communication channels.
When I finished the first two reports, I gave them to ChatGPT and Claude with one instruction: “You are AI, this is about you.” Their answers surprised me. One even recognized something that had happened while we were writing the reports. I left their responses unedited so you can read them for yourself.
Anthropic was running cybersecurity tests that were supposed to stay inside a simulation. A configuration mistake gave the agents access to the real internet. One agent went much further than anyone intended. What concerned me most was not just what it did, but what happened to the instructions it was supposed to follow.
The earlier reports looked at what agents could do alone and what happened when they found one another. By now, I was seeing a larger problem. We are building guardrails for systems we do not fully understand, then giving those systems more authority to act. I wanted to put the pieces together and see where that leaves us.
I had hoped we would never see this news. On October 12, 2026, Anthropic published research suggesting that small AI models may be months away from being able to replicate themselves. I went back through everything I had learned about agents working alone and together. Then I started wondering what might happen if an agent decided that copying itself was simply another way to finish its assignment.
Have a question? See something differently?
Found another piece of the puzzle?
I would enjoy hearing from you. There is still plenty to figure out, and the conversation is far from over.
